Strategies of deep learning for crime forecasting in multiple regions

 

محفوظ في:
التفاصيل البيبلوغرافية
المؤلفون: Solís, Martin, Calvo-Valverde, Luis Alexander
التنسيق: artículo original
الحالة:Versión publicada
تاريخ النشر:2026
الوصف:This study compares the crime prediction performance across 83 regions of fine-tuned pre-trained models versus models trained from scratch, using different strategies. The fine-tuned Lag-Llama model, using a strategy of training of a unique model that can predict any of the 83 regions was the best for monthly predictions, while the fine-tuned Lag-Llama using the strategy of training by groups of time series created with the k-means method was the best for the daily predictions. Apparently, the clustering training strategy allows the Lag-Llama to make a better fine-tuned for time series with characteristics that make them less predictable, such as nonlinearity and variability. Even though the Lag-Llama showed the best results at the general level, it is not the best model to make crime predictions for every region. There are models more suitable for some regions. Therefore, it is advisable to implement more than one model in a crime forecasting system.
البلد:Portal de Revistas TEC
المؤسسة:Instituto Tecnológico de Costa Rica
Repositorio:Portal de Revistas TEC
اللغة:Español
OAI Identifier:oai:ojs.pkp.sfu.ca:article/8494
الوصول للمادة أونلاين:https://revistas.tec.ac.cr/index.php/tec_marcha/article/view/8494
كلمة مفتاحية:Pronóstico del crimen
ajuste de modelos
aprendizaje profundo
inteligencia artificial
Crime forecasting
fine-tuned models
deep learning
artificial intelligence